Catálogo de publicaciones - libros

In this paper, we propose a novel learning method for face detection using discriminative feature selection. The main deficiency of the boosting algorithm for face detection is its long training time. Through statistical learning theory, our discriminative feature selection method can make the training process for face detection much faster than the boosting algorithm without degrading the generalization performance. Being different from the boosting algorithm which works in an iterative learning way, our method can directly solve the learning problem of face detection. Our method is a novel ensemble learning method for combining multiple weak classifiers. The most discriminative component classifiers are selected for the ensemble. Our experiments show that the proposed discriminative feature selection method is more efficient than the boosting algorithm for face detection.

It is very meaningful for dimension reduction by extraction and analysis of the underlying manifold embedded in face observation space, since the low dimensional manifold can represent the varying intrinsic features. However, this kind of manifold is perhaps not useful for face image recognition problem. This paper proposes a new discriminative manifold learning method which can efficiently discover the discriminative manifold. Besides the characteristic of preserving the local structure similarity in the face submanifold, the proposed method emphasizes the discriminative property of embedding much more throughout building and solving an object function. Experimental results on some open face datasets indicate the proposed method can achieve lower error rates.

We propose a novel neural network for classification of visual patterns. The new network, called or PyraNet, has a hierarchical structure with two types of processing layers, namely pyramidal layers and 1-D layers. The PyraNet is motivated by two concepts: the image pyramids and local receptive fields. In the new network, nonlinear 2-D are trained to perform both 2-D analysis and data reduction. In this paper, we present a fast training method for the PyraNet that is based on resilient back-propagation and weight decay, and apply the new network to classify gender from facial images.

In this paper, a novel model for Gabor-based independent radial basis function (IRBF) neural network is proposed and applied to face recognition. In the new model, a bank of Gabor filters is first built to extract Gabor face representations characterized by selected frequency, locality and orientation to cope with various illuminations, facial expression and poses in face recognition. Then principal component analysis (PCA) is adopted to reduce the dimension of the extracted Gabor face representations for every face sample. At last, a new IRBF neural network is built to extract high-order statistical features of extracted Gabor face representations with lower dimension and to classify these extracted high-order statistical features. According to the experiments on the famous CAS-PEAL face database, our proposed approach could outperform ICA with architecture II (ICA2) and kernel PCA (KPCA) with standing testing sets proposed in the current release disk of the CAS-PEAL face database.

In this paper, two approaches to improve the illumination robustness of the face recognition algorithms are presented, that is, Symmetrical Image Correction (SIC) and Bit-Plan Feature Fusion (BPFF). SIC can reduce bright speckles and shadows caused by over lighting. BPFF constructs a new virtual face with Bit-Plan information of face images. Generalized PCA is then applied to the virtual faces to achieve face recognition. Experiments show that, the proposed combined method can reduce the sensitivity of face recognition to illuminations using fewer projection vectors than the compared approaches.

Two-dimensional linear discriminant analysis (2DLDA) was recently developed for face image representation and recognition by adopting the idea of image projection in 2DPCA. 2DLDA outperforms traditional LDA mainly in terms of feature extraction speed. Unfortunately, 2DLDA needs to use large numbers of features to represent an image sample, causing storage requirements are heavy and also feature matching process is time-consuming. Against this problem, we discuss in this paper a new image representation scheme called Enhanced 2DLDA (E-2DLDA) for face recognition. The main strategy adopted in our method is that two image projections are applied to an image sample jointly, so the dimensions of extracted feature matrix along both horizontal direction and vertical direction get compressed, and finally the total number of features can be reduced to a great extent. The experimental results on ORL database show that this method remarkably outperforms existing 2DLDA in terms of speed of feature matching and storage requirements of features.

In this paper, a support vector machine (SVM) based adaptive color switching for human face tracking is proposed. The color space is switching to the most appropriate color space model (CSM) according to circumstance conditions adaptively. Recently, many face tracking algorithms used empirical skin color model to discriminate skin/non-skin regions. These skin color models not consider illumination variation and result in less capacity to model skin color distribution. In this work, four color spaces and Laws texture extracted from face image database are used to train each SVM independently. In the pre-processing, the discrete wavelet transform (DWT) refines the face features would concentrate important features and reduce the computational complexity. Then, the features are transformed into four CSMs for SVMs which provide good generalization through optimal hyperplane. In testing, we perform quality measurement method to evaluate the face tracking performance and aggregating each SVM classification results to color space switching. Experimental results show that the proposed method would switch to the most appropriate color space according to quality measurement, automatically.

A new subspace analysis approach named ANLBM is proposed based on Laplacianfaces. It uses the discriminant information of training samples by supervised mechanism, enhances within-class local information by an objective function. The objective function is used to construct adjacency graph’s weight matrix. In order to avoid the drawback of Laplacianfaces’ PCA step, ANLBM uses kernel mapping. ANLBM changes the problem from minimum eigenvalue solution to maximum eigenvalue solution, reduces the redundancy of the computing and increases the precision of the result. The experiments are performed on ORL and Yale databases. Experimental results show that ANLBM has a better performance.

This article addresses the problem of rotation invariant face detection using convolutional neural networks. Recently, we developed a new class of convolutional neural networks for visual pattern recognition. These networks have a simple network architecture and use shunting inhibitory neurons as the basic computing elements for feature extraction. Three networks with different connection schemes have been developed for in-plane rotation invariant face detection: fully-connected, toeplitz-connected, and binary-connected networks. The three networks are trained using a variant of Levenberg-Marquardt algorithm and tested on a set of 40,000 rotated face patterns. As a face/non-face classifier, these networks achieve 97.3% classification accuracy for a rotation angle in the range ±90 and 95.9% for full in-plane rotation. The proposed networks have fewer free parameters and better generalization ability than the feedforward neural networks, and outperform the conventional convolutional neural networks.

The mean shift algorithm is an efficient technique for object tracking. However, it has a shortcoming that it can’t adjust scale with object during tracking process. There are presently no effective ways to solve this problem. The kernel bandwidth of mean shift tracker in one frame is generally steered by the object scale obtained in the previous frame, so it is very important for mean shift tracker to correctly describe the scale of the target in very frame. In accordance with the kernel-bandwidth effect on the mean shift tracker and the property of face, this paper introduces a new idea that uses direct least square ellipse fitting to adjust the facial scale. The experimental results demonstrate the efficiency of this algorithm. Its performance has been proven superior to the original mean shift tracking algorithm.